Hello! I built mcp-recall to solve the problem of MCP token chewing with Claude Code (also other coding systems). I saw that MCP tool output adds up quickly. A snapshot, a large API pull, a few of those, and half an hour in your context window is full. My session started needed compaction mid-task. What stood out to me: Claude Code already offloads its built-in tools to disk, but MCP tool output just gets cut off at ~25k tokens and dropped. That was the gap I wanted to close. So mcp-recall catches MCP output before it hits the window, keeps the full copy locally in SQLite, and hands Claude a short summary instead. When it needs the details, it pulls back exactly what it needs. There is no need in re-running the tool. Coming from security, keeping it local mattered to me: nothing leaves your machine, no outside service, no real dependencies. It's deterministic too. no second model in the loop deciding what to keep. I built it for my own work, but figured it could be useful for others running long MCP sessions. If you do, I'd like to hear what tends to eat your context the most.

MCP-Recall

Keeps MCP tools' output from filling your context

Keeps MCP tools' output from filling your context